TextEncodeBooguEdit

TextEncodeBooguEdit is the conditioning front-end for Boogu-Image, Alibaba's instruction-editing model - the same family line as Qwen-Image-Edit. You hand it a reference image and a sentence like "turn the background into a rainy street," and it produces the conditioning that makes the model edit the picture instead of just describing it. Where this one differs from the plain Qwen-Edit path is that the reference isn't just a vision prompt - it's also injected as an identity latent, which is the trick that keeps the subject recognizable after the edit.

That dual role is the whole interesting design, and it's a neat piece of CFG engineering worth understanding even if you never touch the node.

How it works

The reference image is used twice, deliberately. First, it goes through the Qwen3-VL vision tower to produce instruction-understanding tokens - those tell the model what to change, and they're attached to the positive conditioning only. Second, it's VAE-encoded into a reference latent representing the image's identity, and that latent goes into both positive and negative. Under classifier-free guidance, something present in both streams cancels out - so the identity survives the CFG subtraction while the instruction (positive-only) gets amplified. Result: the model changes what you asked and keeps the subject looking like itself. If the source-code comment is to be believed, the tokenizer even picks the right system prompt automatically (image-to-image when a reference is present, drop-negatives when the negative is empty), so you don't have to plumb templates by hand.

Inputs that matter

• clip - the Boogu text encoder.
• prompt - the edit instruction.
• negative_prompt - optional; leave empty to let the model drop negatives.
• vae - encodes the reference identity latent.
• images - the reference(s); Boogu focuses on one per sample, more allowed (the schema lets you grow the list).

Outputs: positive and negative CONDITIONING for the sampler.

Where it fits

Ships with ComfyUI core - Boogu-Image support landed June 2026, so this is a new node. The Boogu checkpoint goes in your ComfyUI models folder. In the 2026 landscape, instruction editing has absorbed a huge chunk of what masks used to do, and Boogu slots in beside Qwen-Edit with the notable advantage that the built-in node handles the identity-preservation wiring for you.

Common issues

The same structural caveat as the whole instruction-editing family: the model re-emits the entire frame, so pixels you didn't ask about will move - that's why the standard workflow sometimes bolts a mask back on around these models. Reference quality matters as much as prompt quality: a low-res or badly lit reference caps both the instruction understanding and the identity latent. And "one reference per sample" means throwing four conflicting images in muddies which identity the VAE latent is supposed to lock - keep it to one clean subject image unless you've got a reason.